//---------------------------------------------------------------------------
// @function:
// CLogical::PosFromIndex
//
// @doc:
// Compute an order spec based on an index
//
//---------------------------------------------------------------------------
COrderSpec *
CLogical::PosFromIndex
(
IMemoryPool *pmp,
const IMDIndex *pmdindex,
DrgPcr *pdrgpcr
)
{
// compute the order spec
COrderSpec *pos = GPOS_NEW(pmp) COrderSpec(pmp);
const ULONG ulLenIncludeCols = pmdindex->UlKeys();
for (ULONG ul = 0; ul < ulLenIncludeCols; ul++)
{
ULONG ulPos = pmdindex->UlKey(ul);
CColRef *pcr = (*pdrgpcr)[ulPos];
IMDId *pmdid = pcr->Pmdtype()->PmdidCmp(IMDType::EcmptL);
pmdid->AddRef();
// TODO: March 27th 2012; we hard-code NULL treatment
// need to revisit
pos->Append(pmdid, pcr, COrderSpec::EntLast);
}
return pos;
}
//---------------------------------------------------------------------------
// @function:
// CPhysicalStreamAgg::PosCovering
//
// @doc:
// Construct order spec on grouping column so that it covers required
// order spec, the function returns NULL if no covering order spec
// can be created
//
//---------------------------------------------------------------------------
COrderSpec *
CPhysicalStreamAgg::PosCovering
(
IMemoryPool *mp,
COrderSpec *posRequired,
CColRefArray *pdrgpcrGrp
)
const
{
GPOS_ASSERT(NULL != posRequired);
if (0 == posRequired->UlSortColumns())
{
// required order must be non-empty
return NULL;
}
// create a set of required sort columns
CColRefSet *pcrsReqd = posRequired->PcrsUsed(mp);
COrderSpec *pos = NULL;
CColRefSet *pcrsGrpCols = GPOS_NEW(mp) CColRefSet(mp, pdrgpcrGrp);
if (pcrsGrpCols->ContainsAll(pcrsReqd))
{
// required order columns are included in grouping columns, we can
// construct a covering order spec
pos = GPOS_NEW(mp) COrderSpec(mp);
// extract order expressions from required order
const ULONG ulReqdSortCols = posRequired->UlSortColumns();
for (ULONG ul = 0; ul < ulReqdSortCols; ul++)
{
CColRef *colref = const_cast<CColRef *>(posRequired->Pcr(ul));
IMDId *mdid = posRequired->GetMdIdSortOp(ul);
COrderSpec::ENullTreatment ent = posRequired->Ent(ul);
mdid->AddRef();
pos->Append(mdid, colref, ent);
}
// augment order with remaining grouping columns
const ULONG size = pdrgpcrGrp->Size();
for (ULONG ul = 0; ul < size; ul++)
{
CColRef *colref = (*pdrgpcrGrp)[ul];
if (!pcrsReqd->FMember(colref))
{
IMDId *mdid = colref->RetrieveType()->GetMdidForCmpType(IMDType::EcmptL);
mdid->AddRef();
pos->Append(mdid, colref, COrderSpec::EntLast);
}
}
}
pcrsGrpCols->Release();
pcrsReqd->Release();
return pos;
}
//---------------------------------------------------------------------------
// @function:
// CPhysical::PosDerivePassThruOuter
//
// @doc:
// Helper for common case of sort order derivation
//
//---------------------------------------------------------------------------
COrderSpec *
CPhysical::PosDerivePassThruOuter
(
CExpressionHandle &exprhdl
)
{
COrderSpec *pos = exprhdl.Pdpplan(0 /*child_index*/)->Pos();
pos->AddRef();
return pos;
}
//---------------------------------------------------------------------------
// @function:
// CPhysicalDML::PosComputeRequired
//
// @doc:
// Compute required sort order based on the key information in the table
// descriptor:
// 1. If a table has no keys, no sort order is necessary.
//
// 2. If a table has keys, but they are not modified in the update, no sort
// order is necessary. This relies on the fact that Split always produces
// Delete tuples before Insert tuples, so we cannot have two versions of the
// same tuple on the same time. Consider for example tuple (A: 1, B: 2), where
// A is key and an update "set B=B+1". Since there cannot be any other tuple
// with A=1, and the tuple (1,2) is deleted before tuple (1,3) gets inserted,
// we don't need to enforce specific order of deletes and inserts.
//
// 3. If the update changes a key column, enforce order on the Action column
// to deliver Delete tuples before Insert tuples. This is done to avoid a
// conflict between a newly inserted tuple and an old tuple that is about to be
// deleted. Consider table with tuples (A: 1),(A: 2), where A is key, and
// update "set A=A+1". Split will generate tuples (1,"D"), (2,"I"), (2,"D"), (3,"I").
// If (2,"I") happens before (2,"D") we will have a violation of the key constraint.
// Therefore we need to enforce sort order on Action to get all old tuples
// tuples deleted before the new ones are inserted.
//
//---------------------------------------------------------------------------
COrderSpec *
CPhysicalDML::PosComputeRequired
(
IMemoryPool *pmp,
CTableDescriptor *ptabdesc
)
{
COrderSpec *pos = GPOS_NEW(pmp) COrderSpec(pmp);
const DrgPbs *pdrgpbsKeys = ptabdesc->PdrgpbsKeys();
if (1 < pdrgpbsKeys->UlLength() && CLogicalDML::EdmlUpdate == m_edmlop)
{
// if this is an update on the target table's keys, enforce order on
// the action column, see explanation in function's comment
const ULONG ulKeySets = pdrgpbsKeys->UlLength();
BOOL fNeedsSort = false;
for (ULONG ul = 0; ul < ulKeySets && !fNeedsSort; ul++)
{
CBitSet *pbs = (*pdrgpbsKeys)[ul];
if (!pbs->FDisjoint(m_pbsModified))
{
fNeedsSort = true;
break;
}
}
if (fNeedsSort)
{
IMDId *pmdid = m_pcrAction->Pmdtype()->PmdidCmp(IMDType::EcmptL);
pmdid->AddRef();
pos->Append(pmdid, m_pcrAction, COrderSpec::EntAuto);
}
}
else if (m_ptabdesc->FPartitioned())
{
COptimizerConfig *poconf = COptCtxt::PoctxtFromTLS()->Poconf();
BOOL fInsertSortOnParquet = FInsertSortOnParquet();
BOOL fInsertSortOnRows = FInsertSortOnRows(poconf);
if (fInsertSortOnParquet || fInsertSortOnRows)
{
GPOS_ASSERT(CLogicalDML::EdmlInsert == m_edmlop);
m_fInputSorted = true;
// if this is an INSERT over a partitioned Parquet or Row-oriented table,
// sort tuples by their table oid
IMDId *pmdid = m_pcrTableOid->Pmdtype()->PmdidCmp(IMDType::EcmptL);
pmdid->AddRef();
pos->Append(pmdid, m_pcrTableOid, COrderSpec::EntAuto);
}
}
return pos;
}
//---------------------------------------------------------------------------
// @function:
// CPhysicalStreamAgg::PosRequiredStreamAgg
//
// @doc:
// Compute required sort columns of the n-th child
//
//---------------------------------------------------------------------------
COrderSpec *
CPhysicalStreamAgg::PosRequiredStreamAgg
(
IMemoryPool *mp,
CExpressionHandle &exprhdl,
COrderSpec *posRequired,
ULONG
#ifdef GPOS_DEBUG
child_index
#endif // GPOS_DEBUG
,
CColRefArray *pdrgpcrGrp
)
const
{
GPOS_ASSERT(0 == child_index);
COrderSpec *pos = PosCovering(mp, posRequired, pdrgpcrGrp);
if (NULL == pos)
{
// failed to find a covering order spec, use local order spec
m_pos->AddRef();
pos = m_pos;
}
// extract sort columns from order spec
CColRefSet *pcrs = pos->PcrsUsed(mp);
// get key collection of the relational child
CKeyCollection *pkc = exprhdl.GetRelationalProperties(0)->Pkc();
if (NULL != pkc && pkc->FKey(pcrs, false /*fExactMatch*/))
{
CColRefSet *pcrsReqd = posRequired->PcrsUsed(m_mp);
BOOL fUsesDefinedCols = FUnaryUsesDefinedColumns(pcrsReqd, exprhdl);
pcrsReqd->Release();
if (!fUsesDefinedCols)
{
// we are grouping on child's key,
// stream agg does not need to sort child and we can pass through input spec
pos->Release();
posRequired->AddRef();
pos = posRequired;
}
}
pcrs->Release();
return pos;
}
//---------------------------------------------------------------------------
// @function:
// CPhysicalSequence::PosDerive
//
// @doc:
// Derive sort order
//
//---------------------------------------------------------------------------
COrderSpec *
CPhysicalSequence::PosDerive
(
IMemoryPool *, // pmp,
CExpressionHandle &exprhdl
)
const
{
// pass through sort order from last child
const ULONG ulArity = exprhdl.UlArity();
GPOS_ASSERT(1 <= ulArity);
COrderSpec *pos = exprhdl.Pdpplan(ulArity - 1 /*ulChildIndex*/)->Pos();
pos->AddRef();
return pos;
}
//---------------------------------------------------------------------------
// @function:
// CQueryContext::PqcGenerate
//
// @doc:
// Generate the query context for the given expression and array of
// output column ref ids
//
//---------------------------------------------------------------------------
CQueryContext *
CQueryContext::PqcGenerate
(
IMemoryPool *pmp,
CExpression * pexpr,
DrgPul *pdrgpulQueryOutputColRefId,
DrgPmdname *pdrgpmdname,
BOOL fDeriveStats
)
{
GPOS_ASSERT(NULL != pexpr && NULL != pdrgpulQueryOutputColRefId);
CColRefSet *pcrs = GPOS_NEW(pmp) CColRefSet(pmp);
DrgPcr *pdrgpcr = GPOS_NEW(pmp) DrgPcr(pmp);
COptCtxt *poptctxt = COptCtxt::PoctxtFromTLS();
CColumnFactory *pcf = poptctxt->Pcf();
GPOS_ASSERT(NULL != pcf);
const ULONG ulLen = pdrgpulQueryOutputColRefId->UlLength();
for (ULONG ul = 0; ul < ulLen; ul++)
{
ULONG *pul = (*pdrgpulQueryOutputColRefId)[ul];
GPOS_ASSERT(NULL != pul);
CColRef *pcr = pcf->PcrLookup(*pul);
GPOS_ASSERT(NULL != pcr);
pcrs->Include(pcr);
pdrgpcr->Append(pcr);
}
COrderSpec *pos = NULL;
CExpression *pexprResult = pexpr;
COperator *popTop = PopTop(pexpr);
if (COperator::EopLogicalLimit == popTop->Eopid())
{
// top level operator is a limit, copy order spec to query context
pos = CLogicalLimit::PopConvert(popTop)->Pos();
pos->AddRef();
}
else
{
// no order required
pos = GPOS_NEW(pmp) COrderSpec(pmp);
}
CDistributionSpec *pds = NULL;
BOOL fDML = CUtils::FLogicalDML(pexpr->Pop());
poptctxt->MarkDMLQuery(fDML);
if (fDML)
{
pds = GPOS_NEW(pmp) CDistributionSpecAny(COperator::EopSentinel);
}
else
{
pds = GPOS_NEW(pmp) CDistributionSpecSingleton(CDistributionSpecSingleton::EstMaster);
}
CRewindabilitySpec *prs = GPOS_NEW(pmp) CRewindabilitySpec(CRewindabilitySpec::ErtNone /*ert*/);
CEnfdOrder *peo = GPOS_NEW(pmp) CEnfdOrder(pos, CEnfdOrder::EomSatisfy);
// we require satisfy matching on distribution since final query results must be sent to master
CEnfdDistribution *ped = GPOS_NEW(pmp) CEnfdDistribution(pds, CEnfdDistribution::EdmSatisfy);
CEnfdRewindability *per = GPOS_NEW(pmp) CEnfdRewindability(prs, CEnfdRewindability::ErmSatisfy);
CCTEReq *pcter = poptctxt->Pcteinfo()->PcterProducers(pmp);
CReqdPropPlan *prpp = GPOS_NEW(pmp) CReqdPropPlan(pcrs, peo, ped, per, pcter);
pdrgpmdname->AddRef();
return GPOS_NEW(pmp) CQueryContext(pmp, pexprResult, prpp, pdrgpcr, pdrgpmdname, fDeriveStats);
}
//---------------------------------------------------------------------------
// @function:
// CQueryContext::PqcGenerate
//
// @doc:
// Generate the query context for the given expression and array of
// output column ref ids
//
//---------------------------------------------------------------------------
CQueryContext *
CQueryContext::PqcGenerate
(
IMemoryPool *mp,
CExpression * pexpr,
ULongPtrArray *pdrgpulQueryOutputColRefId,
CMDNameArray *pdrgpmdname,
BOOL fDeriveStats
)
{
GPOS_ASSERT(NULL != pexpr && NULL != pdrgpulQueryOutputColRefId);
CColRefSet *pcrs = GPOS_NEW(mp) CColRefSet(mp);
CColRefArray *colref_array = GPOS_NEW(mp) CColRefArray(mp);
COptCtxt *poptctxt = COptCtxt::PoctxtFromTLS();
CColumnFactory *col_factory = poptctxt->Pcf();
GPOS_ASSERT(NULL != col_factory);
// Collect required column references (colref_array)
const ULONG length = pdrgpulQueryOutputColRefId->Size();
for (ULONG ul = 0; ul < length; ul++)
{
ULONG *pul = (*pdrgpulQueryOutputColRefId)[ul];
GPOS_ASSERT(NULL != pul);
CColRef *colref = col_factory->LookupColRef(*pul);
GPOS_ASSERT(NULL != colref);
pcrs->Include(colref);
colref_array->Append(colref);
}
// Collect required properties (prpp) at the top level:
// By default no sort order requirement is added, unless the root operator in
// the input logical expression is a LIMIT. This is because Orca always
// attaches top level Sort to a LIMIT node.
COrderSpec *pos = NULL;
CExpression *pexprResult = pexpr;
COperator *popTop = PopTop(pexpr);
if (COperator::EopLogicalLimit == popTop->Eopid())
{
// top level operator is a limit, copy order spec to query context
pos = CLogicalLimit::PopConvert(popTop)->Pos();
pos->AddRef();
}
else
{
// no order required
pos = GPOS_NEW(mp) COrderSpec(mp);
}
CDistributionSpec *pds = NULL;
BOOL fDML = CUtils::FLogicalDML(pexpr->Pop());
poptctxt->MarkDMLQuery(fDML);
// DML commands do not have distribution requirement. Otherwise the
// distribution requirement is Singleton.
if (fDML)
{
pds = GPOS_NEW(mp) CDistributionSpecAny(COperator::EopSentinel);
}
else
{
pds = GPOS_NEW(mp) CDistributionSpecSingleton(CDistributionSpecSingleton::EstMaster);
}
// By default, no rewindability requirement needs to be satisfied at the top level
CRewindabilitySpec *prs = GPOS_NEW(mp) CRewindabilitySpec(CRewindabilitySpec::ErtNotRewindable, CRewindabilitySpec::EmhtNoMotion);
// Ensure order, distribution and rewindability meet 'satisfy' matching at the top level
CEnfdOrder *peo = GPOS_NEW(mp) CEnfdOrder(pos, CEnfdOrder::EomSatisfy);
CEnfdDistribution *ped = GPOS_NEW(mp) CEnfdDistribution(pds, CEnfdDistribution::EdmSatisfy);
CEnfdRewindability *per = GPOS_NEW(mp) CEnfdRewindability(prs, CEnfdRewindability::ErmSatisfy);
// Required CTEs are obtained from the CTEInfo global information in the optimizer context
CCTEReq *pcter = poptctxt->Pcteinfo()->PcterProducers(mp);
// NB: Partition propagation requirements are not initialized here. They are
// constructed later based on derived relation properties (CPartInfo) by
// CReqdPropPlan::InitReqdPartitionPropagation().
CReqdPropPlan *prpp = GPOS_NEW(mp) CReqdPropPlan(pcrs, peo, ped, per, pcter);
// Finally, create the CQueryContext
pdrgpmdname->AddRef();
return GPOS_NEW(mp) CQueryContext(mp, pexprResult, prpp, colref_array, pdrgpmdname, fDeriveStats);
}
//---------------------------------------------------------------------------
// @function:
// CPhysicalSequenceProject::CreateOrderSpec
//
// @doc:
// Create local order spec that we request relational child to satisfy
//
//---------------------------------------------------------------------------
void
CPhysicalSequenceProject::CreateOrderSpec
(
IMemoryPool *pmp
)
{
GPOS_ASSERT(NULL == m_pos);
GPOS_ASSERT(NULL != m_pds);
GPOS_ASSERT(NULL != m_pdrgpos);
m_pos = GPOS_NEW(pmp) COrderSpec(pmp);
// add partition by keys to order spec
if (CDistributionSpec::EdtHashed == m_pds->Edt())
{
CDistributionSpecHashed *pdshashed = CDistributionSpecHashed::PdsConvert(m_pds);
const DrgPexpr *pdrgpexpr = pdshashed->Pdrgpexpr();
const ULONG ulSize = pdrgpexpr->UlLength();
for (ULONG ul = 0; ul < ulSize; ul++)
{
CExpression *pexpr = (*pdrgpexpr)[ul];
// we assume partition-by keys are always scalar idents
CScalarIdent *popScId = CScalarIdent::PopConvert(pexpr->Pop());
const CColRef *pcr = popScId->Pcr();
gpmd::IMDId *pmdid = pcr->Pmdtype()->PmdidCmp(IMDType::EcmptL);
pmdid->AddRef();
m_pos->Append(pmdid, pcr, COrderSpec::EntLast);
}
}
if (0 == m_pdrgpos->UlLength())
{
return;
}
COrderSpec *posFirst = (*m_pdrgpos)[0];
#ifdef GPOS_DEBUG
const ULONG ulLength = m_pdrgpos->UlLength();
for (ULONG ul = 1; ul < ulLength; ul++)
{
COrderSpec *posCurrent = (*m_pdrgpos)[ul];
GPOS_ASSERT(posFirst->FSatisfies(posCurrent) &&
"first order spec must satisfy all other order specs");
}
#endif // GPOS_DEBUG
// we assume here that the first order spec in the children array satisfies all other
// order specs in the array, this happens as part of the initial normalization
// so we need to add columns only from the first order spec
const ULONG ulSize = posFirst->UlSortColumns();
for (ULONG ul = 0; ul < ulSize; ul++)
{
const CColRef *pcr = posFirst->Pcr(ul);
gpmd::IMDId *pmdid = posFirst->PmdidSortOp(ul);
pmdid->AddRef();
COrderSpec::ENullTreatment ent = posFirst->Ent(ul);
m_pos->Append(pmdid, pcr, ent);
}
}